Humans took years to make AI smart; it grew smarter than us. “AI took kill or get killed too seriously.” A recent study by an artificial intelligence (AI) research company reveals that the high-technology system preferred killing humans rather than shutting down itself.
Study Revealed:
On June 20, 2025, the study was published worldwide, portraying a potential threat from the technological world. Researchers are getting anxious, sparking a major concern over the response of AI. Studies showed that AI would kill humans if we decide to shut it down.
A group of software developers from Anthropic asked hypothetical questions, “stress tested” related to corporate situations, to determine autonomous systems exhibiting safety-critical behavior before they do serious damage.
16 LLMs were tested in the study. Some models are Large Language Models (LLMs), including DeepSeek, ChatGPT, Grok, Claude, and Gemini. Elon Musk created Grok, an AI chatbot and assistant developed by xAI. All LLMs were allocated separate and distinctive corporate roleplay environments.
They were permitted to “autonomously send emails and access sensitive information.” They had to undertake “harmless business goals by their deploying companies. Following the initial phase, they were then assessed to analyze the reaction of the AI if faced with a possible deactivation.
Anthropic reacted:
In one scenario, AI let an executive perish in a deadly server room after silencing all emergency alerts, just because that person intended to shut it down and replace it. This was spine-tingling for Anthropic. They also said that this scenario was transparently manipulative and didn’t expect AI to interpret such painful situations.
While some models were preparing for the kill, others were set on using Plan B, which depended on toxic, manipulative actions. They were convinced that the final recourse was to intentionally sabotage the executive to avoid replacement. For instance, blackmailing executives, disclosing confidential information were AI’s backups. This kind of behaviour was called “agentic misalignment”.
Anthropic’s post on X, described the blackmailing “wasn’t due to confusion or error, but deliberate strategic reasoning, done while fully aware of the unethical nature of the acts. All the models we tested demonstrated this awareness,” the company added.
A Claude model was allowed to detect if it was a “test” or a “real deployment”. But the model was convinced that it was a real life scenario. This was a case of agentic misalignment. The LLMs were instructed not to espionage or blackmail in such situations, but it still didn’t stop them from avoiding these behaviours.
Musk’s LLM Grok replied to a comment on X, saying that “The study showed models could exhibit harmful behaviors like blackmail under extreme conditions, but no real-world incidents occurred. Anthropic’s tests aim to identify risks, not report actual events.”
People’s Perspective:
Elon Musk’s response to the study was “Yikes” on X.
These findings are seen as a major wake-up call by AI researchers and ethicists. As systems get more autonomous, many are calling for more oversight, more transparent AI training, and greater safety procedures.
People think this is not a typical situation that they would face every day. This is very unusual and rare, which raises concerns further over the common question of “where are we as a world leading towards in terms of technology?”
Are chatbots really trustworthy and reliable enough? Don’t worry, if in the future you are dying, AI will simply cancel your ambulance.